Recurrent Sinusoidal INRs for Efficient High-Fidelity Representation

This paper studies sinusoidal recurrence as an iterative mechanism for harmonic spectral enrichment in implicit neural representations (INRs). It argues that sinusoidal activations generate a harmonic line spectrum, while recurrent unrolling progressively expands the effective spectral support. The authors implement this idea with a shared sinusoidal block that repeatedly refines a latent representation. According to the abstract, the method achieves higher fidelity than feed-forward INR baselines on RGB image benchmarks with fewer parameters and optimization steps, and transfers favorably to super-resolution, NeRF, and SDF tasks. The supplied summary does not provide detailed benchmark numbers or ablation results.